“…In addition to predicting reaction outcomes for substrates not seen by the model, it would be advantageous to be able to predict reactivity for ligands that the model was not trained on and identify ideal ligand candidates for Ni-catalyzed borylation. A proximity-based ligand space guided search to select 12 new ligands (Figure , red circles) for studying the ML models trained on the original 24 ligands (Figure , blue circles) was performed . The selected ligand set includes eight ligands expected to perform well based on Kraken values (CM-Phos, CPhos, CX-POMeCy, Cy-JohnPhos-OMe, PPh 2 ( o -Anis), PPhCy 2 , RuPhos, and S-Phos) and four ligands expected to perform poorly (PEt 3 , PBn 3 , CX-PInCy, and P( p -CF 3 -Ph) 3 ).…”